Documentation
agentrc is organized as a small set of stable documents rather than a monolithic application manual.
What is agentrc?
Start here: a Dockerfile-shaped recipe for portable, governed agents, the problem it solves, and why it is needed.
Specification
The full 0.1.0-draft.6 working draft: the four new keywords, the /mnt layout, build → ai.agentrc.* labels, and Cedar enforcement.
Quickstart
Write your first Agentfile, build it (BuildKit frontend or arc build), read the emitted labels, push, and run.
Demo: one agent, three backends
Build the code-reviewer agent once, then run the same artifact on three substrates — local microVM, Bedrock, and Kubernetes — via the reference translators.
Agentfile
The single-agent declaration: IDENTITY, CAPABILITY, SOP, POLICY over standard Dockerfile keywords.
Security
POLICY requests become ai.agentrc.* labels; the platform grants/narrows/rejects and enforces via Cedar, deny-by-default.
Package model
A standard OCI artifact: layers carry /mnt resources, the config carries ai.agentrc.* labels. Sign, mirror, review before run.
Runners
How a platform reads ai.agentrc.* labels (never the Agentfile) and a substrate executes CMD — without agentrc becoming a runtime.
Conformance
Profile-based conformance and the adversarial test suite that probes the fail-closed, deny-by-default guarantees a platform must uphold.
Implementation mapping
How today's reference-implementation modules line up with the 0.1.0-draft.6 Agentfile model, gaps called out honestly.
Reference implementation
The Go tooling/ that implements both build paths — try arc lint, the BuildKit frontend, and the emitted ai.agentrc.* labels.